How every credit report and KYC dossier is built, graded and sized. The same in every market.
Each report is a policy simulation: the output of Avexo's underwriting and KYC agents run on one listed company, using public data only. It shows what the infrastructure produces. It is not a loan offer, a sanction, a credit rating or investment advice, and Avexo does not lend.
Every company gets two documents. The credit report asks whether the company can be lent to and on what terms. The KYC and integrity dossier asks who the entity is, who controls it and what the public record says about its conduct.
| Market | Statements | Notes and disclosures | Registers |
|---|---|---|---|
| India | Financial statements as filed with the exchanges, standardised | Read from the annual report by a language model, then cross-checked against the statements | Exchange shareholding filings; aggregator company records |
| United States | Form 10-K and 10-Q on SEC EDGAR, read by hand; market data for revenue and EBITDA | The 10-K notes, read by hand and recorded | SEC filer record; 13F, 13D, 13G and Form 4 filings; the proxy statement |
| United Kingdom | Audited consolidated accounts and the interim report, read by hand | The notes to the same accounts | Companies House officers, persons with significant control, charges and filings |
Every calculation is deterministic: the same recorded inputs always produce the same document. A language model may read a note or phrase a sentence; it never computes a number or decides a pass or fail.
Two axes are scored from 1 (strongest) to 4 (weakest). Financial risk reads leverage, interest cover, cash conversion, total outside liabilities to net worth (after guarantees where they exist), operating cash before its largest single swing, and free cash flow. Business risk reads the business model, how capital splits across segments, revenue billed to related parties, conduct toward small suppliers where the law requires it to be disclosed, and how much of the file public sources could not reach.
The two scores give an anchor letter. Published notches then move it for what the axes cannot see: a promoter pledge or non-disposal undertaking on the exchange record, two or more blocks of the file not available publicly, or a material acquisition completed after the statements. Notches are capped at two down and one up.
A fired covenant caps the grade at C, because a breached threshold is an open question by definition. Fired covenants do not also notch, since most are already scored inside an axis and would otherwise count twice.
The grade runs A to E and is a policy action, not a rating. It carries no probability of default: there is no default study behind it. The same letter carries three actions, one per risk appetite, because the grade states the risk and the lender's appetite decides what to do about it.
Twelve or thirteen covenants, depending on the market, are tested against the latest year. Thresholds depend on the business model, because 135 debtor days is alarming for a distributor and normal for a project contractor. Each covenant reports its distance to the threshold, including those that have not fired, and any covenant within 10% of its threshold is marked near.
Limits come from the same policy engine that sizes a live application. It builds a limit from two paths: a working-capital path, which funds a share of the gap between current assets and the current liabilities already financing them, and an earnings path, which sizes against gross profit and operating cash. A fixed rule combines the two and a policy ceiling applies.
The engine runs at three standing appetites. A conservative appetite holds back more of the working-capital gap as owner margin, funds a smaller share of what remains and has a lower ceiling; a growth appetite does the opposite and buys the difference with controls. The ceilings are USD 500,000, 1,000,000 and 2,000,000, converted at a stated, dated rate outside the US. The bands are fixed policy, so the spread between them describes our appetite, never the company.
How much of the engine runs depends on the filing. UK accounts and US 10-Ks give the current and non-current split and separate cost of sales from operating cost, so both paths run. India's standardised statements carry neither, so only the monthly gross-profit leg runs and the rest is reported as not computable rather than estimated. Every report shows the model-implied figure beside the ceiling, so a limit set by policy is never mistaken for a view of the company.
Every detected condition, a fired covenant, a disclosure gap or a note-level finding, produces a question, a control and a monitoring line. Conditions are grouped into a handful of risks. The wording of each risk, question, control and monitor is a fixed template keyed on which conditions fired, never written for the company, and every line traces to the figure that raised it. The register is the decision; the grade is its index.
Where a figure appears in two independent places, the two are compared at a 2% tolerance: a note against the statements, market data against the filing, or the copy filed at a registry against the published report. A figure implied from a ratio is an estimate: it can be in line with a reported balance or off it, but it cannot corroborate it. A disagreement is published, never reconciled away, and a figure that fails as an extraction defect is withheld from the assessment.
Each report ends by listing what first-party data would have settled in that specific case. In India that means GST returns, bank statements and the credit bureaus; in the US and UK, bank statements, the sales and purchase ledgers, a lien or charge search and the bureau. The same feeds drive our network model, which maps a borrower's customers, suppliers and related parties and reads concentration and contagion across them.